With advances in technology, large amounts of data are generated regularly. This data is derived from real-world applications, such as text classification, spam detection, and medical and engineering fields. This massive amount of data contains a feature set of various sizes. Not all of these features are required for discovering and analyzing knowledge. Most of these features might be irrelevant or redundant to the problem. These unwanted features are not useful for handling classification problems. They decrease the computational simplicity and increase the classification error [1]. One of the significant preprocessing techniques is dimensionality reduction (DR). The purpose of DR is to reduce the number of features while meeting certain criteria and achieving superior performance. Feature Selection (FS) is one of the most substantial tools in DR. It is the process of choosing the most relevant and effective features to decrease classification error and manage the dimensionality of the data set by discarding unnecessary features [2].

There are four main stages of FS [3]. The first stage is the creation of a feature subset; the second stage is the assessment of the feature subset; the third stage is the establishment of termination criteria; and the fourth stage is validation. During the initial phase, various methods are employed to generate a feature subset using a search strategy. In the second stage, an objective function, also known as a fitness function and a classifier, is used to assess the quality of produced solutions. This procedure is repeated until the stopping criteria are reached. A wrapper-based technique is one of the FS approaches. It uses a specific classifier to find optimal or near-optimal solutions. The objective of FS (Feature Selection) is to find the optimum subset of features. It can be considered as a search procedure. Calculating 2N−1 subsets of a dataset with N features is an NP-hard problem, where NP-hard refers to a class of problems for which no efficient solution algorithm is known. The best subset cannot be obtained using an exhaustive search algorithm (which tries all possible combinations).

Nature-inspired algorithms have validated strong performance compared to classical optimization methods, mainly for nonlinear and complex real-world problems. According to the No Free Lunch (NFL) theorem [4], no single optimization algorithm can outperform all others for every problem. This motivates the ongoing development and compositing of efficacy from known procedures. Hybridizing functions allows the incorporation of complementary strengths from multiple algorithms to be better balanced between exploitation and exploration. The Flower Pollination Algorithm (FPA) [5] works fine for global optimization but can suffer from slow convergence due to a stochastic Lévy flight global search. The Sine Cosine Algorithm (SCA) [6] demonstrates durable exploration, but is limited in exploitation when calibrating into a local node search. Thus, combining the exploration capabilities of SCA and the exploitation of FPA can create a more balanced and efficient search procedure. The FPA component adds a layer to enhance local and global pollination dynamics. Together, these two phases sequentially coordinate to have the algorithm converge quickly. Thus, the design enables the proposed model to achieve higher classification accuracy while obtaining fewer but essential features across the multiple benchmark datasets. In FS, the datasets are typically high-dimensional, with idle features, and many complex interactions among features. These characteristics create difficulty for a single algorithm in achieving effective exploratory and efficient exploitative characteristics. Given these practical challenges, the developed hSC-FPA is inspired to find a balance created for a specific problem. The main contributions of this paper are as follows: 1.

A new hybridized method, hSC-FPA, supported by the NFL nature of nature-inspired algorithms and tailored for FS problems, is proposed.

The rest of the paper is organized as follows: Section 2 presents a literature review of the hybridization of metaheuristic algorithms. Section 3 discusses the preliminaries. In Section 4, the methodology of the hybrid model is discussed in detail. Section 5 provides insights into the experimental setup. In Section 6, numerical experimentation and performance comparison of the hSC-FPA with other algorithms are presented. Finally, Section 7 concludes the paper and suggests some future directions.

Mafarja and Mirjalili [7] proposed a hybrid method, HBALO (Hybrid Binary Ant Lion Optimizer), integrating the Binary Ant Lion Optimizer with two incremental hill-climbing techniques, QuickReduct and CEBARKCC, to manage the complexity of FS. A novel hybrid FS algorithm called Mayfly-Harmony Search (MA-HS), combining the Mayfly Algorithm with Harmony Search to improve search space exploitation, was introduced by Bhattacharyya et al. [8]. Results showed the superior performance of MA-HS in selecting informative features and improving classification outcomes. Abualigah and Dulaimi [9] introduced the hybrid Sine Cosine Algorithm and Genetic Algorithm. The algorithm was tested on UCI (University of California, Irvine) datasets using metrics such as classification accuracy, fitness values, and the number of selected features. Mohammadzadeh and Gharehchopogh [10] proposed a hybrid firefly singed Sine Cosine Flower Pollination Algorithm called HWOAFPA. In this process, the Whale Optimization Algorithm was enhanced with opposition-based learning, and its final populations were used to initialize the flower pollination algorithm with improved convergence speed and accuracy. Sun et al. [11] presented a Feature Selection technique known as MetaSCA, which is based on Sine Cosine Algorithm (SCA), but enhances the original SCA by introducing features like the golden sine section coefficient and multi-level tuning factor. Abdelhamid et al. [12] introduced a new binary hybrid metaheuristic algorithm for FS called bSCWDTO, which combined the Sine Cosine Algorithm and Dipper Throated Optimization. Ragab [13] proposed a hybrid binary metaheuristic approach that integrated Particle Swarm Optimization and Firefly Algorithm. It was designed to optimize classification accuracy while minimizing the number of features for FS in machine learning classification problems. Hammouri et al. [14] proposed three improved binary versions of the White Shark Optimizer as a method for FS to address challenges with pre-mature convergence and local optima. The binary versions, Binary Distribution-based White Shark Optimizer (BDWSO), Binary Sine Cosine White Shark Optimizer (BSCWSO), and Binary Hybrid Sine Cosine White Shark Optimizer (BHSCWSO), introduced improvements for FS, including distribution-based updates, adaptive functions from sine-cosine algorithms, and hybrid acceleration strategies. Alkhonaini et al. [15] proposed HSCCOFS-DL, a hybrid deep learning-based method for threat recognition in IoT sensor networks. HSCCOFS-DL employed the Hybrid Sine-Cosine Chimp Optimization Algorithm to perform FS, used a symmetrical autoencoder to perform a classification task, and deployed the Sparrow Search Algorithm to tune the hyperparameters.

Kaur et al. [16] looked forward to applying an automated method to determine quality in Golden Delicious apples with an image processing and feature extraction method. Kaur was able to extract a total of 18,654 features from 1256 images and needed a technique for feature selection. For feature selection, Minimum Redundancy Maximum Relevance was used. Kumari et al. [17] addressed the shortcomings of conventional intrusion detection systems by presenting a hybrid model that successfully combined the Flower Pollination Algorithm, Cheetah Optimization Algorithm, and Artificial Neural Networks to increase accuracy and lower false alarms. Azeem et al. [18] presented a new combination of PSO-HHO to promote FS, advancing the strengths of both metaheuristics to eliminate unnecessary features while improving their classifier performance. A wrapper-based approach using K-Nearest Neighbors (K-NN) with Euclidean distance was employed to evaluate candidate solutions. Aly et al. [19] presented a hybrid metaheuristic algorithm HB-GWO (Hybrid Butterfly-Grey Wolf Optimizer), which combined the Butterfly Optimization Algorithm (BOA) and the Grey Wolf Optimizer (GWO) for effective feature selection in high-dimensional datasets. By using an adaptive switching mechanism to balance exploration and exploitation, HB-GWO achieves superior performance across benchmark datasets when evaluated with a Random Forest classifier. Experimental studies showed that HB-GWO obtained 96.8%, 80.6%, 89.5%, and 72.4% accuracy on breast cancer, colon cancer, and arrhythmia datasets, respectively.

Most existing hybrid metaheuristics emphasize convergence or exploration but overlook the balance between accuracy and feature reduction. To address these gaps, the proposed hSC-FPA enhances the balance between exploration and exploitation, confirming stable convergence and improved accuracy across various datasets.

This sections provides the FS problem design and a summary of the conventional Sine Cosine Algorithm (SCA) and Flower Pollination Algorithm (FPA).

The suggested hybrid hSC-FPA for feature selection is presented here along with its three main steps: initialization, binary conversion, and hybrid optimization. In Algorithm 1, step-by-step instructions together with the underlying mathematics for the algorithm are presented.

The efficiency of the suggested hSC-FPA algorithm is evaluated using several measurement criteria, including average accuracy and average select size. Classical SCA and FPA are tested along with the proposed hSC-FPA. Furthermore, a comparative analysis was conducted with other hybrid models already proposed in the literature for FS. The proposed optimizer employs a hybrid approach to obtain the optimal solution, which involves maximizing accuracy while maintaining an optimal subset of features.

The following Table 3 presents the classification accuracies of FPA, SCA, and the proposed hybrid hSC-FPA across 14 benchmark datasets. In datasets such as CNAE, Ionosphere, Sonar, and Wine, the hybrid model surpasses FPA and SCA. This shows improved generality and FS capability. On data like Iris, Soybean Small, Ecoli, hSC-FPA performs competitively. The gains are especially pronounced in more complex datasets, such as Sonar and Tic-Tac-Toe, where hSC-FPA shows greater improvements.

Table 4 compares the feature selection sizes from FPA, SCA, and the hybrid hSC-FPA across benchmark datasets. SCA usually generates the fewest features in most datasets. For datasets like Iris, Soybean Small, and Wine, hSC-FPA gives the least number of features. Although hSC-FPA does not always yield the smallest amount, it often avoids selecting too many features, especially in comparison to FPA. This indicates that the hybrid method is effective for both precision and for getting small, useful feature subsets. Thus, the goal of achieving a Pareto-optimal solution is met, as the proposed hSC-FPA reaches high classification accuracy while balancing accuracy and feature reduction effectively.

The Wilcoxon’s rank-sum test was performed at the 0.05 significance level to study the statistical difference between the proposed method and the alternative feature selection methods, FPA and SCA. When comparing the proposed algorithm’s output to other algorithms, this test can help reveal whether or not the results differ significantly. The proposed algorithm’s findings differ considerably from those of the compared methods when the p-value is less than 0.05. On the other hand, insignificant results are indicated by a p-value of more than 0.05. The worst p-values in Table 5 are those greater than 0.05, which is the significance level applied. When associating the proposed method with SCA and FPA, the table shows that the p-values obtained with this test are less than 0.05 across datasets such as CNAE, Sonar, Spam base, Tic-tac-toe, and waveform. The results prove the statistical significance of the hSC-FPA method.

The variation in performance of the proposed hSC-FPA algorithm across datasets can be attributed to differences in dataset dimensionality and structure. For high-dimensional datasets such as CNAE and Spam Base, the hybridization mechanism supports effective navigation of a large and sparse feature space. This results in higher accuracy and a substantial reduction in irrelevant features. Conversely, for smaller or low-dimensional datasets such as Iris and Zoo, the search landscape is less complex, and therefore, the advantage of hybrid feature selection is less pronounced. As a result, the performance of hSC-FPA becomes comparable to that of other algorithms, since the reduced dimensionality requires less exploration effort and allows competing methods to achieve similar convergence and classification outcomes. This behavior aligns with the convergence patterns shown in Fig. 2, where high-dimensional datasets exhibit wider initial fluctuations due to exploration, while low-dimensional datasets converge rapidly with minimal variance across iterations.

Table 6 provides a comparative evaluation of the proposed hSC-FPA algorithm relative to other existing hybrid optimization algorithms from the literature.

Implementing feature selection (FS) before the learning stage is crucial for improving classification performance. This approach simplifies the model. This work presents a hybrid method, hSC-FPA, which combines the exploration strengths of SCA with the exploitation abilities of FPA. We use the K-NN classifier for evaluation. In the proposed algorithm, SCA’s global search behavior boosts population diversity, while FPA ensures effective local refinement. The algorithm was tested on fourteen benchmark datasets to evaluate its accuracy and FS effectiveness. Its performance was compared with traditional FPA and SCA. Although the hSC-FPA shows good performance across various benchmark datasets, there are some limitations to consider. As a population-based metaheuristic algorithm, it can sometimes become stuck in local optima, particularly in high-dimensional or complex search spaces. Additionally, hybridizing SCA and FPA adds computational overhead when using single-algorithm strategies, especially with large datasets. Future research will aim to address these limitations by creating adaptive or parallel implementations of hSC-FPA to reduce computation time without losing accuracy. Since the algorithm’s main strength lies in balancing exploration and exploitation, future extensions could explore dynamic parameter control to further enhance this balance. Experiments demonstrated higher classification accuracy with compact feature subsets. As future directions, we suggest utilizing the proposed hybrid optimizer to address other real-world problems, including scheduling issues, engineering optimization problems, and molecular potential energy functions. The presented methodology can be applied to other popular classifiers, such as Artificial Neural Network (ANN), Random Forest, and Support Vector Machine (SVM), and examined to determine whether performance is stable or varies. Another potential future work is to hybridize the FPA with other machine learning and deep learning algorithms.